Raw LLM Responses

7:05 What Chris Olah is saying is that the AI, when presented with a dataset, will form connections between differently formatted pieces of data. The method that it uses to do this, while deterministic at its core, is happening on such a scale that it _seems_ random to a human observer. Thus, Chris Olah could not tell you _exactly_ how it works. He could, however, tell you what I just did, albeit to a far greater degree of specificity since I only have a surface-level understanding of how LLMs work. This is a similar issue to what happened with NFTs a few years ago - there is a massive disconnect between what the thing actually _does_ and what people _think_ it does, and it's created largely out of how the people selling it to the general public are going about presenting it. The difference is that, while the folks hawking NFTs are doing so with payouts in the millions, Generative AI is something that some of the largest companies in the world (Google, Microsoft, Amazon) have spent _hundreds of billions of dollars_ pursuing. They therefore have a much greater vested interest in justifying that expense to the market at large, and one of the ways they do it is by dramatically overselling how complex the AI actually is and what it is actually capable of. By all means, feel free to call out these companies for their total disregard for the safety of their product. However, I would also encourage asking as a follow-up: "is it more likely that these people are genuinely pursuing something that can actively decide to kill its creators, or that they know that it can't, but think that making it sound like they are is a really convincing sales pitch for how super-smart their chatbot is?"